3 Modelling Simple Toxicity Endpoints …
45
KEs in the AOP unless they were de-prioritised by the exclusion criteria) to predict the
hazard (sensitiser/non-sensitiser). The potency of chemicals assigned as sensitisers
is then predicted using a similar mechanistic read across process within a structural
alert as described previously, except now considering both known human skin and
mouse skin sensitisation data for analogue chemicals (Fig. 3.2).
The DA accurately predicted skin sensitisation hazard when compared against
both LLNA data and human data (85% DA vs. LLNA; 86% DA vs. human) indicating
that the DA can predict both species equally well. Furthermore, the DA is more
predictive of human skin sensitisation (86% accuracy—DA vs. human) than the
LLNA (81% accuracy—LLNA vs. human) for the dataset analysed (Table 3.3).
The DA correctly predicted the Basketter potency category (5 categories: 1–4, 5/6)
for 59% and the GHS classification (three categories: 1A, 1B, no category) for 73%
of the evaluation dataset when compared to LLNA data. When compared against
human data, the DA correctly predicted the Basketter potency category for 68% and
the GHS classification for 76% of the dataset, respectively (Table 3.4). The DA was
also shown to predict both classification methods for human data more accurately
than the LLNA (DA vs. human, 68% Basketter and 76% GHS; LLNA vs. human,
54% Basketter and 65% GHS).
3.5 Enabling Expert Review
Whilst in silico tools can provide accurate predictions in isolation [24, 25], expert
review adds significant value to the accuracy of the conclusions drawn [26–29].
There has been considerable effort to delineate the types of information that should
be considered when carrying out expert review of in silico predictions, particularly
for the endpoint of mutagenicity [24, 30–33].
Knowledge of inadequacies in the test system being modelled as well as the arguments associated with the data used to generate the prediction should be considered
when carrying out expert review. Analysis of the results presented, bearing in mind
strengths and limitations associated with different modelling techniques, is also an
important factor in the assessment of the predictions generated.
Therefore, it is important that any prediction systems employed in this context
provide enough detailed information about how the prediction was derived so that
they can be probed by the expert user to support the overall decision they make. Some
general considerations which should be made during the expert review process and
the information that should be provided by the predictive system to allow for this
analysis are:
1. Limitations of the test being modelled
2. Relevance and adequacy of the data used to make a prediction
3. Similarity of the query compound to the compounds used to make the prediction
4. Coverage of any potential toxicophores
5. Causality of any toxicophore identified
45
KEs in the AOP unless they were de-prioritised by the exclusion criteria) to predict the
hazard (sensitiser/non-sensitiser). The potency of chemicals assigned as sensitisers
is then predicted using a similar mechanistic read across process within a structural
alert as described previously, except now considering both known human skin and
mouse skin sensitisation data for analogue chemicals (Fig. 3.2).
The DA accurately predicted skin sensitisation hazard when compared against
both LLNA data and human data (85% DA vs. LLNA; 86% DA vs. human) indicating
that the DA can predict both species equally well. Furthermore, the DA is more
predictive of human skin sensitisation (86% accuracy—DA vs. human) than the
LLNA (81% accuracy—LLNA vs. human) for the dataset analysed (Table 3.3).
The DA correctly predicted the Basketter potency category (5 categories: 1–4, 5/6)
for 59% and the GHS classification (three categories: 1A, 1B, no category) for 73%
of the evaluation dataset when compared to LLNA data. When compared against
human data, the DA correctly predicted the Basketter potency category for 68% and
the GHS classification for 76% of the dataset, respectively (Table 3.4). The DA was
also shown to predict both classification methods for human data more accurately
than the LLNA (DA vs. human, 68% Basketter and 76% GHS; LLNA vs. human,
54% Basketter and 65% GHS).
3.5 Enabling Expert Review
Whilst in silico tools can provide accurate predictions in isolation [24, 25], expert
review adds significant value to the accuracy of the conclusions drawn [26–29].
There has been considerable effort to delineate the types of information that should
be considered when carrying out expert review of in silico predictions, particularly
for the endpoint of mutagenicity [24, 30–33].
Knowledge of inadequacies in the test system being modelled as well as the arguments associated with the data used to generate the prediction should be considered
when carrying out expert review. Analysis of the results presented, bearing in mind
strengths and limitations associated with different modelling techniques, is also an
important factor in the assessment of the predictions generated.
Therefore, it is important that any prediction systems employed in this context
provide enough detailed information about how the prediction was derived so that
they can be probed by the expert user to support the overall decision they make. Some
general considerations which should be made during the expert review process and
the information that should be provided by the predictive system to allow for this
analysis are:
1. Limitations of the test being modelled
2. Relevance and adequacy of the data used to make a prediction
3. Similarity of the query compound to the compounds used to make the prediction
4. Coverage of any potential toxicophores
5. Causality of any toxicophore identified